Re: Some functions are not inlined by gcc 3.2, resulting code is ugly

[Denis Vlasenko]

On 3 November 2002 14:17, Jussi Laako wrote:
> On Sun, 2002-11-03 at 18:17, Denis Vlasenko wrote:
>
> Jump target 17e0 is aligned (with nops):
> >     17dd:	88 02                	mov    %al,(%edx)
> >     17df:	90                   	nop
> >     17e0:	89 d0                	mov    %edx,%eax
> >     17e2:	5a                   	pop    %edx
> >
> >     17ec:	eb f2                	jmp    17e0
> > <__constant_memcpy+0x20>
> >
> >     17fa:	eb e4                	jmp    17e0
> > <__constant_memcpy+0x20>
> >
> >     1800:	eb de                	jmp    17e0
> > <__constant_memcpy+0x20>
> >
> >     187c:	e9 5f ff ff ff       	jmp    17e0
> > <__constant_memcpy+0x20> 1881:	eb 0d                	jmp    1890
> > <__constant_memcpy+0xd0> 1883:	90                   	nop
>
> ...
>
> >     188f:	90                   	nop
> >     1890:	c1 e9 02             	shr    $0x2,%ecx
> >     1893:	89 d7                	mov    %edx,%edi
>
> And also jump target 1890 is aligned.
>
>
> I think the penalty of few NOPs is smaller than result of jump to
> unaligned address. This is especially true with P4 architecture.

Alignment does not eliminate jump. It only moves jump target to 16 byte
boundary. This _probably_ makes execution slightly faster but on average
it costs you 7,5 bytes. This price is too high when you take into account
L1 instruction cache wastage and current bus/core clock ratios.
--
vda